| As an important part of national development strategy,autonomous driving technology has significant development potential and economic benefits.As one of the important components of the automatic driving system,the environment perception module can collect information about the surrounding environment autonomously,calculate the distribution of obstacles around the vehicle,the vehicle’s own position and a map of the vehicle’s vicinity,which in turn provides the decision basis for path planning.Vehicles need to make decisions quickly in order to cope with unexpected situations in the process of moving,so autonomous driving systems have strict limits on the complexity and inference time of the environment perception algorithms.However,the automatic driving system needs to synthesize the prediction results of multiple tasks to make correct travel decisions,and using multiple independent task processing units inevitably will bring a great challenge to computational resources and real-time performance.Therefore,exploring a multi-task learning method to build deep learning models that can simultaneously complete multiple tasks such as target detection,instance segmentation,and target tracking to reduce time overhead and computational requirements is an important problem that needs to be solved in the field of automatic driving today.In this thesis,we propose a environment perception model based on deep learning for simultaneous multi-task learning of three tasks,including target detection,instance segmentation and target tracking,to address the needs of autonomous driving systems for multi-task scenarios.For the problem of lack of intensive scene data in the public dataset,a data acquisition platform is built to expand the public dataset.Meanwhile,a decoupling-fusion module based on attention mechanism is constructed to improve the multi-task learning model,which improves the detection accuracy of the model for the three tasks.In addition,the self-collected dataset is used for transfer learning of the model to improve the generalization ability of the model.The main work of the thesis is as follows:(1)Aiming at the problem of slow prediction speed of detection algorithms for multiple independent tasks,the multi-task learning method is introduced and a multi-task environment perception model is proposed to simultaneously complete three tasks: target detection,instance segmentation and target tracking.Aiming at the problem that three tasks interfere with each other and cause knowledge confusion during the learning process,a decoupling-fusion module based on an attention mechanism is built to alleviate the phenomenon of negative transfer between tasks,so as to improve the detection accuracy of the algorithm on each task.Aiming at the problem of the training conflicting among tasks and fails to converge,the dynamic weighted average method is used to train the model and dynamically adjust the decline rate of the loss function of each task,so as to achieve the balance of the convergence speed of different tasks.(2)Aiming at the problem that most of the current public datasets are for urban areas,highways and suburban areas at home and abroad which have open views and low obstacle density.A data collection platform is built and data collection,data pre-processing,data annotation and data analysis in dense scenes such as universities and parks.The public datasets is expanded.(3)The improved algorithm is tested on the KITTI dataset and the transfer learning is carried out on the self-collected dataset,the ablation experiment is designed to verify the effectiveness of the modules of the improved algorithm finally.The advantages of the improved algorithm in detecting far and small targets,detecting occluded and truncated targets,segmentation mask extraction and target trajectory reconstruction are also verified through visual comparison analysis. |
